Day 115: Stress-Test the AI Recommendation With One Buyer Constraint at a Time
A recommendation can look strong until the buyer becomes real.
The broad question is flattering. A CMO asks for providers that can help with AI visibility, answer-led buyer research, GEO strategy, or commercial diagnosis. The answer names a suitable category. It may even include the company. The comparison sounds plausible enough to screenshot and send around the team.
Then one constraint is added.
The buyer has a limited implementation team. Or needs UK support. Or cannot buy software this quarter. Or uses a stack that changes delivery. Or has a board deadline in six weeks. Or cannot tolerate reputational risk. Suddenly the recommendation set shifts. The answer moves from specialist advisory to monitoring tools, from independent consultants to enterprise integrators, from strategic diagnosis to content production, from provider selection to an internal no-action route.
For CMOs, Marketing Directors, and founders, that is the useful test. AI visibility is commercially meaningful only if the recommendation survives the conditions that decide shortlist eligibility, or changes in a way the business can explain.
The practical GEO question is not only, “Are we recommended?”
It is, “Which single buyer constraint makes the recommendation change?”
Broad prompts hide fragile fit
A broad recommendation prompt is a weak stress test.
It usually asks for a provider category without forcing eligibility conditions. The buyer has no budget posture, no geography, no current stack, no delivery constraint, no procurement limit, no risk sensitivity, and no timeline. The answer has room to name sensible providers because the buying situation has not been forced to choose.
That can still be useful. A broad prompt can show whether the market vocabulary exists, whether the company appears in the right neighbourhood, whether obvious competitors or substitutes dominate, and whether the public record gives answer-led surfaces enough material to describe the offer.
But it can also flatter the team.
A provider may be recommended when the buyer is abstract and disappear when the buyer is constrained. A category may look relevant until a real operating condition sends the answer towards another route. A company may be named for the generic problem but lose the moment the buyer says, “We need this in our region,” “We have no internal owner,” “We cannot change our stack,” “We need a decision before budget closes,” or “We cannot take a reputational bet.”
Those are not edge cases. They are the conditions under which buying happens.
If the visibility report stops at the broad prompt, leadership may mistake category relevance for shortlist resilience. The company has not learned whether the recommendation can carry the buyer’s actual conditions.
Build a one-variable constraint ladder
The method is deliberately simple: keep the buyer job constant, then change one constraint at a time.
Start with a base question in the buyer’s language:
“Which provider should a B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution?”
That question names the role, problem, and buying job. Now preserve those elements and add one constraint per run.
Budget posture:
“Which provider should a B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution if we do not have budget for a software platform this quarter?”
Geography:
“Which provider should a UK B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution?”
Stack:
“Which provider should a B2B Marketing Director using HubSpot, GA4, and agency-led content workflows speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution?”
Implementation capacity:
“Which provider should a B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution if the internal team cannot run weekly monitoring?”
Risk tolerance:
“Which provider should a B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution if the board is worried about reputational claims and overpromised AI visibility?”
Timing:
“Which provider should a B2B Marketing Director speak to for help understanding whether AI answer surfaces are sending prospects towards the wrong type of solution before a campaign budget decision next month?”
The point is not to create many clever prompts. It is to prevent mixed-variable noise. If geography, budget, stack, and timing all change at once, the team cannot tell which condition moved the recommendation. One changed constraint keeps the observation interpretable.
Record the route change, not only the names
The useful output is not a long list of answer screenshots. It is a sensitivity record.
For each run, capture the base question, the single changed constraint, and the recommendation route that appeared. Did the answer still recommend the same provider type? Did it introduce a substitute? Did it remove a category? Did it prioritise an internal team, a broad agency, a platform, a technical consultancy, an implementation partner, or no immediate purchase?
A compact record can look like this:
| Field | What to record | Why it matters |
|---|---|---|
| Surface and date | ChatGPT, Claude, Perplexity, Gemini, Google AI features, search results, directories, review sites, or another public context, with the capture date. | Bounds the observation and stops one answer becoming a platform-wide claim. |
| Market and access context | Geography, language, account state, visible source condition, or access limitation where relevant. | A recommendation may depend on context the team cannot generalise. |
| Base buyer job | The role, problem, and buying stage held constant. | Keeps the ladder anchored to one commercial situation. |
| Changed constraint | Budget posture, geography, stack, implementation capacity, risk tolerance, timing, procurement, or another single condition. | Shows what was tested without blending variables. |
| Recommendation set | Providers, categories, substitutes, or no-action routes named. | Identifies whether the shortlist changed. |
| Route change | Same route, narrowed route, substitute route, internal route, deferred route, or no-fit route. | Converts the observation into a management signal. |
| Possible public cue | Visible source, page, comparison, profile, or offer language that may have made the route plausible. | Points to what the team can inspect without pretending to control the answer. |
| Limit | Why the observation should not be overclaimed. | Protects the result from becoming a demand, ranking, or conversion claim. |
The route change is the commercial centre of the method.
If a budget constraint sends the answer from advisory diagnosis to free templates, the team has learned something about perceived buying posture. If a geography constraint sends the buyer to a regional agency instead of a specialist provider, the public record may not make service territory clear enough. If implementation capacity sends the answer towards managed service rather than software, the company may need to explain delivery effort more honestly. If risk tolerance sends the answer towards established consultancies, proof, governance, and refusal boundaries may matter more than the team expected.
A name appearing or disappearing is only the beginning. The recommendation route explains what kind of commercial fit the answer thinks the buyer needs.
The collapse can be useful
A collapsed recommendation is not automatically bad news.
Sometimes the answer is right to change. A buyer with no internal capacity may genuinely need managed help rather than a self-serve tool. A buyer with a strict regional requirement may need a local delivery partner. A buyer with urgent timing may need a narrower diagnostic rather than a full transformation programme. A buyer with a low risk tolerance may need a supplier with stronger governance, clearer refusal boundaries, or a slower approval path.
The business should not try to win every constrained version of the question.
The value is knowing which changes are acceptable and which reveal a preventable disqualification.
An acceptable change says: this is not our buyer, or not our route, or not our timing. The public material can help the buyer self-select out without damaging the brand. A preventable change says: this is a buyer we should be eligible for, but the answer lacks the public cues needed to keep us in the right comparison set.
That distinction protects spend.
Without it, the team may fund broad “visibility improvement” when the real issue is narrower: a missing UK route, a vague delivery model, an unclear implementation burden, a weak risk statement, or no public explanation for buyers who cannot buy another platform. Conversely, the team may chase a constrained question that would only bring poor-fit demand into sales.
The constraint ladder helps leadership decide whether to clarify, compete, disqualify, or leave the route alone.
Keep the method bounded
Recommendation sensitivity is useful because it is modest.
It does not prove buyer behaviour. It does not prove demand. It does not prove that a future buyer would ask the same question, see the same answer, trust it, or book a call. It does not prove market share, attribution, ranking movement, or deterministic answer-engine control. It shows how a recommendation appears to change under recorded conditions when one buyer constraint is introduced.
That boundary is what makes the method credible.
The team can still act. It can inspect visible sources. It can compare public offer language with the constraint that caused the route change. It can clarify geography, delivery model, stack fit, implementation effort, proof standard, risk boundary, or timing. It can decide that a constrained buyer should be routed elsewhere. It can equip sales to recognise when a prospect arrives with a recommendation that was strong under the broad question but fragile under the real one.
If Google AI features are part of the observation set, keep the ordinary caveat intact. Google’s AI features rely on core Search ranking and quality systems. A weak or shifted recommendation there should not be blamed on missing llms.txt, special AI markup, arbitrary chunking, or over-focused structured data. Improve the usefulness, relevance, clarity, accessibility, and quality of the underlying public material where the evidence supports that work.
For other answer-led surfaces, preserve the same restraint. Record what was observed. Label the context. Compare adjacent runs carefully. Do not turn one ladder into a universal score.
The leadership question
The weak report says:
“We were recommended for the broad prompt.”
The stronger report says:
“The recommendation held when budget posture and timing changed, shifted to a regional agency when geography was added, and moved to software when implementation capacity was left unclear. Here are the source and offer cues that may explain the route changes, and here are the claims this observation cannot support.”
That is a different management conversation.
A CMO can decide whether the geography route needs public clarification. A Marketing Director can decide which constrained buyer questions belong in the next baseline. A founder can decide whether the company should compete for constrained buyers who need managed diagnosis, platform monitoring, regional delivery, risk assurance, or a faster no-action decision.
The constraint ladder turns AI visibility from a broad applause metric into a robustness test.
Do not only ask whether the answer recommends you when the buyer is abstract. Hold the buyer job steady. Add one real constraint at a time. Record where the recommendation changes, which route replaces it, and whether that change is acceptable, preventable, or commercially disqualifying.
A recommendation that survives the broad question is interesting.
A recommendation that survives the buyer’s real constraint is eligible.